v1

llm/generate/gen_common.sh

+# common logic across linux and darwin
+# 
+
+init_vars() {
+    case "${GOARCH}" in
+    "amd64")
+        ARCH="x86_64"
+        ;;
+    "arm64")
+        ARCH="arm64"
+        ;;
+    *)  
+        ARCH=$(uname -m | sed -e "s/aarch64/arm64/g")
+    esac
+
+    LLAMACPP_DIR=../llama.cpp
+    CMAKE_DEFS=""
+    # 这里和llama.cpp不一样
+    CMAKE_TARGETS="--target ollama_llama_server"
+    if echo "${CGO_CFLAGS}" | grep -- '-g' >/dev/null; then
+        CMAKE_DEFS="-DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_GPROF=on -DLLAMA_SERVER_VERBOSE=on ${CMAKE_DEFS}"
+    else
+        # TODO - add additional optimization flags...
+        CMAKE_DEFS="-DCMAKE_BUILD_TYPE=Release -DLLAMA_SERVER_VERBOSE=off ${CMAKE_DEFS}"
+    fi
+    case $(uname -s) in
+    "Darwin")
+        LIB_EXT="dylib"
+        WHOLE_ARCHIVE="-Wl,-force_load"
+        NO_WHOLE_ARCHIVE=""
+        GCC_ARCH="-arch ${ARCH}"
+        ;;
+    "Linux")
+        LIB_EXT="so"
+        WHOLE_ARCHIVE="-Wl,--whole-archive"
+        NO_WHOLE_ARCHIVE="-Wl,--no-whole-archive"
+
+        # Cross compiling not supported on linux - Use docker
+        GCC_ARCH=""
+        ;;
+    *)
+        ;;
+    esac
+    if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
+        CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
+    fi
+}
+
+git_module_setup() {
+    if [ -n "${OLLAMA_SKIP_PATCHING}" ]; then
+        echo "Skipping submodule initialization"
+        return
+    fi
+    # Make sure the tree is clean after the directory moves
+    if [ -d "${LLAMACPP_DIR}/gguf" ]; then
+        echo "Cleaning up old submodule"
+        rm -rf ${LLAMACPP_DIR}
+    fi
+    # git submodule init
+    # git submodule update --force ${LLAMACPP_DIR}
+
+}
+
+apply_patches() {
+    # Wire up our CMakefile
+    if ! grep ollama ${LLAMACPP_DIR}/CMakeLists.txt; then
+        echo 'add_subdirectory(../ext_server ext_server) # ollama' >>${LLAMACPP_DIR}/CMakeLists.txt
+    fi
+
+    if [ -n "$(ls -A ../patches/*.diff)" ]; then
+        # apply temporary patches until fix is upstream
+        for patch in ../patches/*.diff; do
+            for file in $(grep "^+++ " ${patch} | cut -f2 -d' ' | cut -f2- -d/); do
+                (cd ${LLAMACPP_DIR}; git checkout ${file})
+            done
+        done
+        for patch in ../patches/*.diff; do
+            (cd ${LLAMACPP_DIR} && git apply ${patch})
+        done
+    fi
+}
+
+build() {
+    cmake -S ${LLAMACPP_DIR} -B ${BUILD_DIR} ${CMAKE_DEFS}
+    cmake --build ${BUILD_DIR} ${CMAKE_TARGETS} -j8
+}
+
+compress() {
+    echo "Compressing payloads to reduce overall binary size..."
+    pids=""
+    rm -rf ${BUILD_DIR}/bin/*.gz
+    for f in ${BUILD_DIR}/bin/* ; do
+        gzip -n --best -f ${f} &
+        pids+=" $!"
+    done
+    # check for lib directory
+    if [ -d ${BUILD_DIR}/lib ]; then
+        for f in ${BUILD_DIR}/lib/* ; do
+            gzip -n --best -f ${f} &
+            pids+=" $!"
+        done
+    fi
+    echo
+    for pid in ${pids}; do
+        wait $pid
+    done
+    echo "Finished compression"
+}
+
+# Keep the local tree clean after we're done with the build
+cleanup() {
+    (cd ${LLAMACPP_DIR}/ && git checkout CMakeLists.txt)
+
+    if [ -n "$(ls -A ../patches/*.diff)" ]; then
+        for patch in ../patches/*.diff; do
+            for file in $(grep "^+++ " ${patch} | cut -f2 -d' ' | cut -f2- -d/); do
+                (cd ${LLAMACPP_DIR}; git checkout ${file})
+            done
+        done
+    fi
+}

llm/generate/gen_darwin.sh

+#!/bin/bash
+# This script is intended to run inside the go generate
+# working directory must be ./llm/generate/
+
+# TODO - add hardening to detect missing tools (cmake, etc.)
+
+set -ex
+set -o pipefail
+echo "Starting darwin generate script"
+source $(dirname $0)/gen_common.sh
+init_vars
+git_module_setup
+apply_patches
+
+sign() {
+    if [ -n "$APPLE_IDENTITY" ]; then
+        codesign -f --timestamp --deep --options=runtime --sign "$APPLE_IDENTITY" --identifier ai.ollama.ollama $1
+    fi
+}
+
+COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DLLAMA_METAL_EMBED_LIBRARY=on"
+
+case "${GOARCH}" in
+"amd64")
+    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
+
+    # Static build for linking into the Go binary
+    init_vars
+    CMAKE_TARGETS="--target llama --target ggml"
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    BUILD_DIR="../build/darwin/${ARCH}_static"
+    echo "Building static library"
+    build
+
+    if [ -z "$OLLAMA_SKIP_CPU_GENERATE" ]; then
+        #
+        # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
+        #
+        init_vars
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        BUILD_DIR="../build/darwin/${ARCH}/cpu"
+        echo "Building LCD CPU"
+        build
+        sign ${BUILD_DIR}/bin/ollama_llama_server
+        compress
+
+        #
+        # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
+        # Approximately 400% faster than LCD on same CPU
+        #
+        init_vars
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
+        echo "Building AVX CPU"
+        build
+        sign ${BUILD_DIR}/bin/ollama_llama_server
+        compress
+
+        #
+        # ~2013 CPU Dynamic library
+        # Approximately 10% faster than AVX on same CPU
+        #
+        init_vars
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
+        echo "Building AVX2 CPU"
+        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
+        build
+        sign ${BUILD_DIR}/bin/ollama_llama_server
+        compress
+    fi
+    ;;
+"arm64")
+
+    # Static build for linking into the Go binary
+    init_vars
+    CMAKE_TARGETS="--target llama --target ggml"
+    CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DBUILD_SHARED_LIBS=off -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    BUILD_DIR="../build/darwin/${ARCH}_static"
+    echo "Building static library"
+    build
+
+    if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
+        init_vars
+        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
+        BUILD_DIR="../build/darwin/${ARCH}/metal"
+        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
+        build
+        sign ${BUILD_DIR}/bin/ollama_llama_server
+        compress
+    fi
+    ;;
+*)
+    echo "GOARCH must be set"
+    echo "this script is meant to be run from within go generate"
+    exit 1
+    ;;
+esac
+
+cleanup
+echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"

llm/generate/gen_linux.sh

+#!/bin/bash
+# This script is intended to run inside the go generate
+# working directory must be llm/generate/
+
+# First we build one or more CPU based LLM libraries
+#
+# Then if we detect CUDA, we build a CUDA dynamic library, and carry the required
+# library dependencies
+#
+# Then if we detect ROCm, we build a dynamically loaded ROCm lib.  The ROCM
+# libraries are quite large, and also dynamically load data files at runtime
+# which in turn are large, so we don't attempt to cary them as payload
+
+set -ex
+set -o pipefail
+
+# See https://llvm.org/docs/AMDGPUUsage.html#processors for reference
+amdGPUs() {
+    if [ -n "${AMDGPU_TARGETS}" ]; then
+        echo "${AMDGPU_TARGETS}"
+        return
+    fi
+    GPU_LIST=(
+        "gfx900"
+        "gfx906:xnack-"
+        "gfx908:xnack-"
+        "gfx90a:xnack+"
+        "gfx90a:xnack-"
+        "gfx906"
+        "gfx928"
+        "gfx940"
+        "gfx941"
+        "gfx942"
+        "gfx1010"
+        "gfx1012"
+        "gfx1030"
+        "gfx1100"
+        "gfx1101"
+        "gfx1102"
+    )
+    (
+        IFS=$';'
+        echo "'${GPU_LIST[*]}'"
+    )
+}
+
+echo "Starting linux generate script"
+if [ -z "${CUDACXX}" ]; then
+    if [ -x /usr/local/cuda/bin/nvcc ]; then
+        export CUDACXX=/usr/local/cuda/bin/nvcc
+    else
+        # Try the default location in case it exists
+        export CUDACXX=$(command -v nvcc)
+    fi
+fi
+COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off"
+source $(dirname $0)/gen_common.sh
+init_vars
+# git_module_setup
+apply_patches
+
+init_vars
+if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ]; then
+    # Builds by default, allows skipping, forces build if OLLAMA_CPU_TARGET="static"
+    # Enables optimized Dockerfile builds using a blanket skip and targeted overrides
+    # Static build for linking into the Go binary
+    init_vars
+    CMAKE_TARGETS="--target llama --target ggml"
+    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    BUILD_DIR="../build/linux/${ARCH}_static"
+    echo "Building static library"
+    build
+fi
+
+init_vars
+if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
+    # Users building from source can tune the exact flags we pass to cmake for configuring
+    # llama.cpp, and we'll build only 1 CPU variant in that case as the default.
+    if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
+        init_vars
+        echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        BUILD_DIR="../build/linux/${ARCH}/cpu"
+        echo "Building custom CPU"
+        build
+        compress
+    else
+        # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
+        # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+        # -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
+        # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+        # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+        # Note: the following seem to yield slower results than AVX2 - ymmv
+        # -DLLAMA_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
+        # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
+        # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake
+
+        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off"
+        if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
+            #
+            # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
+            #
+            init_vars
+            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+            BUILD_DIR="../build/linux/${ARCH}/cpu"
+            echo "Building LCD CPU"
+            build
+            compress
+        fi
+
+        if [ "${ARCH}" == "x86_64" ]; then
+            #
+            # ARM chips in M1/M2/M3-based MACs and NVidia Tegra devices do not currently support avx extensions.
+            #
+            if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx" ]; then
+                #
+                # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
+                # Approximately 400% faster than LCD on same CPU
+                #
+                init_vars
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+                BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
+                echo "Building AVX CPU"
+                build
+                compress
+            fi
+
+            if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx2" ]; then
+                #
+                # ~2013 CPU Dynamic library
+                # Approximately 10% faster than AVX on same CPU
+                #
+                init_vars
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+                BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
+                echo "Building AVX2 CPU"
+                build
+                compress
+            fi
+        fi
+    fi
+else
+    echo "Skipping CPU generation step as requested"
+fi
+
+# If needed, look for the default CUDA toolkit location
+if [ -z "${CUDA_LIB_DIR}" ] && [ -d /usr/local/cuda/lib64 ]; then
+    CUDA_LIB_DIR=/usr/local/cuda/lib64
+fi
+
+# If needed, look for CUDA on Arch Linux
+if [ -z "${CUDA_LIB_DIR}" ] && [ -d /opt/cuda/targets/x86_64-linux/lib ]; then
+    CUDA_LIB_DIR=/opt/cuda/targets/x86_64-linux/lib
+fi
+
+# Allow override in case libcudart is in the wrong place
+if [ -z "${CUDART_LIB_DIR}" ]; then
+    CUDART_LIB_DIR="${CUDA_LIB_DIR}"
+fi
+
+if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
+    echo "CUDA libraries detected - building dynamic CUDA library"
+    init_vars
+    CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
+    if [ -n "${CUDA_MAJOR}" ]; then
+        CUDA_VARIANT=_v${CUDA_MAJOR}
+    fi
+    if [ "${ARCH}" == "arm64" ]; then
+        echo "ARM CPU detected - disabling unsupported AVX instructions"
+
+        # ARM-based CPUs such as M1 and Tegra do not support AVX extensions.
+        #
+        # CUDA compute < 6.0 lacks proper FP16 support on ARM.
+        # Disabling has minimal performance effect while maintaining compatibility.
+        ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
+    fi
+    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
+    if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
+        echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
+        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
+        echo "Building custom CUDA GPU"
+    else
+        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
+    fi
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
+    BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
+    EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
+    build
+
+    # Carry the CUDA libs as payloads to help reduce dependency burden on users
+    #
+    # TODO - in the future we may shift to packaging these separately and conditionally
+    #        downloading them in the install script.
+    DEPS="$(ldd ${BUILD_DIR}/bin/ollama_llama_server )"
+    for lib in libcudart.so libcublas.so libcublasLt.so ; do
+        DEP=$(echo "${DEPS}" | grep ${lib} | cut -f1 -d' ' | xargs || true)
+        if [ -n "${DEP}" -a -e "${CUDA_LIB_DIR}/${DEP}" ]; then
+            cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/bin/"
+        elif [ -e "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" ]; then
+            cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/bin/"
+        elif [ -e "${CUDART_LIB_DIR}/${lib}" ]; then
+            cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/bin/"
+        else
+            cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/bin/"
+        fi
+    done
+    compress
+
+fi
+
+if [ -z "${ONEAPI_ROOT}" ]; then
+    # Try the default location in case it exists
+    ONEAPI_ROOT=/opt/intel/oneapi
+fi
+
+if [ -d "${ONEAPI_ROOT}" ]; then
+    echo "OneAPI libraries detected - building dynamic OneAPI library"
+    init_vars
+    source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
+    CC=icx
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL=ON -DLLAMA_SYCL_F16=OFF"
+    BUILD_DIR="../build/linux/${ARCH}/oneapi"
+    EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
+    DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
+    build
+
+    # copy oneAPI dependencies
+    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e sycl -e mkl -e tbb); do
+        cp "${dep}" "${BUILD_DIR}/bin/"
+    done
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${BUILD_DIR}/bin/"
+    compress
+fi
+
+if [ -z "${ROCM_PATH}" ]; then
+    # Try the default location in case it exists
+    ROCM_PATH=/opt/rocm
+fi
+
+if [ -z "${CLBlast_DIR}" ]; then
+    # Try the default location in case it exists
+    if [ -d /usr/lib/cmake/CLBlast ]; then
+        export CLBlast_DIR=/usr/lib/cmake/CLBlast
+    fi
+fi
+
+if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
+    echo "ROCm libraries detected - building dynamic ROCm library"
+    if [ -f ${ROCM_PATH}/lib/librocblas.so.*.*.????? ]; then
+        ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
+    fi
+    init_vars
+    echo $(amdGPUs)
+
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs)"
+    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
+    if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
+        echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
+        CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
+        echo "Building custom ROCM GPU"
+    fi
+    BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
+    EXTRA_LIBS="-L${ROCM_PATH}/lib -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
+    build
+
+    # Record the ROCM dependencies
+    rm -f "${BUILD_DIR}/bin/deps.txt"
+    touch "${BUILD_DIR}/bin/deps.txt"
+    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
+        echo "${dep}" >> "${BUILD_DIR}/bin/deps.txt"
+    done
+    # bomb out if for some reason we didn't get a few deps
+    #if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
+    #    cat "${BUILD_DIR}/bin/deps.txt"
+    #    echo "ERROR: deps file short"
+    #    exit 1
+    #fi
+    # compress
+fi
+
+cleanup
+echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"

llm/generate/gen_windows.ps1

+#!powershell
+
+$ErrorActionPreference = "Stop"
+
+function amdGPUs {
+    if ($env:AMDGPU_TARGETS) {
+        return $env:AMDGPU_TARGETS
+    }
+    # TODO - load from some common data file for linux + windows build consistency
+    $GPU_LIST = @(
+        "gfx900"
+        "gfx906:xnack-"
+        "gfx908:xnack-"
+        "gfx90a:xnack+"
+        "gfx90a:xnack-"
+        "gfx940"
+        "gfx941"
+        "gfx942"
+        "gfx1010"
+        "gfx1012"
+        "gfx1030"
+        "gfx1100"
+        "gfx1101"
+        "gfx1102"
+    )
+    $GPU_LIST -join ';'
+}
+
+
+function init_vars {
+    if (!$script:SRC_DIR) {
+        $script:SRC_DIR = $(resolve-path "..\..\")
+    }
+    if (!$script:llamacppDir) {
+        $script:llamacppDir = "../llama.cpp"
+    }
+    if (!$script:cmakeTargets) {
+        $script:cmakeTargets = @("ollama_llama_server")
+    }
+    $script:cmakeDefs = @(
+        "-DBUILD_SHARED_LIBS=on",
+        "-DLLAMA_NATIVE=off"
+        )
+    $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
+    $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
+    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\ollama_runners"
+    md "$script:DIST_BASE" -ea 0 > $null
+    if ($env:CGO_CFLAGS -contains "-g") {
+        $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
+        $script:config = "RelWithDebInfo"
+    } else {
+        $script:cmakeDefs += @("-DLLAMA_SERVER_VERBOSE=off", "-DCMAKE_BUILD_TYPE=Release")
+        $script:config = "Release"
+    }
+    if ($null -ne $env:CMAKE_SYSTEM_VERSION) {
+        $script:cmakeDefs += @("-DCMAKE_SYSTEM_VERSION=${env:CMAKE_SYSTEM_VERSION}")
+    }
+    # Try to find the CUDA dir
+    if ($env:CUDA_LIB_DIR -eq $null) {
+        $d=(get-command -ea 'silentlycontinue' nvcc).path
+        if ($d -ne $null) {
+            $script:CUDA_LIB_DIR=($d| split-path -parent)
+            $script:CUDA_INCLUDE_DIR=($script:CUDA_LIB_DIR|split-path -parent)+"\include"
+        }
+    } else {
+        $script:CUDA_LIB_DIR=$env:CUDA_LIB_DIR
+    }
+    $script:DUMPBIN=(get-command -ea 'silentlycontinue' dumpbin).path
+    if ($null -eq $env:CMAKE_CUDA_ARCHITECTURES) {
+        $script:CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
+    } else {
+        $script:CMAKE_CUDA_ARCHITECTURES=$env:CMAKE_CUDA_ARCHITECTURES
+    }
+    # Note: Windows Kits 10 signtool crashes with GCP's plugin
+    if ($null -eq $env:SIGN_TOOL) {
+        ${script:SignTool}="C:\Program Files (x86)\Windows Kits\8.1\bin\x64\signtool.exe"
+    } else {
+        ${script:SignTool}=${env:SIGN_TOOL}
+    }
+    if ("${env:KEY_CONTAINER}") {
+        ${script:OLLAMA_CERT}=$(resolve-path "${script:SRC_DIR}\ollama_inc.crt")
+    }
+}
+
+function git_module_setup {
+    # TODO add flags to skip the init/patch logic to make it easier to mod llama.cpp code in-repo
+    & git submodule init
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+    & git submodule update --force "${script:llamacppDir}"
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+}
+
+function apply_patches {
+    # Wire up our CMakefile
+    if (!(Select-String -Path "${script:llamacppDir}/CMakeLists.txt" -Pattern 'ollama')) {
+        Add-Content -Path "${script:llamacppDir}/CMakeLists.txt" -Value 'add_subdirectory(../ext_server ext_server) # ollama'
+    }
+
+    # Apply temporary patches until fix is upstream
+    $patches = Get-ChildItem "../patches/*.diff"
+    foreach ($patch in $patches) {
+        # Extract file paths from the patch file
+        $filePaths = Get-Content $patch.FullName | Where-Object { $_ -match '^\+\+\+ ' } | ForEach-Object {
+            $parts = $_ -split ' '
+            ($parts[1] -split '/', 2)[1]
+        }
+
+        # Checkout each file
+        foreach ($file in $filePaths) {
+            git -C "${script:llamacppDir}" checkout $file
+        }
+    }
+
+    # Apply each patch
+    foreach ($patch in $patches) {
+        git -C "${script:llamacppDir}" apply $patch.FullName
+    }
+}
+
+function build {
+    write-host "generating config with: cmake -S ${script:llamacppDir} -B $script:buildDir $script:cmakeDefs"
+    & cmake --version
+    & cmake -S "${script:llamacppDir}" -B $script:buildDir $script:cmakeDefs
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ })"
+    & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ })
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+    # Rearrange output to be consistent between different generators
+    if ($null -ne ${script:config} -And (test-path -path "${script:buildDir}/bin/${script:config}" ) ) {
+        mv -force "${script:buildDir}/bin/${script:config}/*" "${script:buildDir}/bin/"
+        remove-item "${script:buildDir}/bin/${script:config}"
+    }
+}
+
+function sign {
+    if ("${env:KEY_CONTAINER}") {
+        write-host "Signing ${script:buildDir}/bin/*.exe  ${script:buildDir}/bin/*.dll"
+        foreach ($file in @(get-childitem "${script:buildDir}/bin/*.exe") + @(get-childitem "${script:buildDir}/bin/*.dll")){
+            & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
+                /csp "Google Cloud KMS Provider" /kc "${env:KEY_CONTAINER}" $file
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
+    }
+}
+
+function install {
+    write-host "Installing binaries to dist dir ${script:distDir}"
+    mkdir ${script:distDir} -ErrorAction SilentlyContinue
+    $binaries = dir "${script:buildDir}/bin/*.exe"
+    foreach ($file in $binaries) {
+        copy-item -Path $file -Destination ${script:distDir} -Force
+    }
+
+    write-host "Installing dlls to dist dir ${script:distDir}"
+    $dlls = dir "${script:buildDir}/bin/*.dll"
+    foreach ($file in $dlls) {
+        copy-item -Path $file -Destination ${script:distDir} -Force
+    }
+}
+
+function cleanup {
+    $patches = Get-ChildItem "../patches/*.diff"
+    foreach ($patch in $patches) {
+        # Extract file paths from the patch file
+        $filePaths = Get-Content $patch.FullName | Where-Object { $_ -match '^\+\+\+ ' } | ForEach-Object {
+            $parts = $_ -split ' '
+            ($parts[1] -split '/', 2)[1]
+        }
+
+        # Checkout each file
+        foreach ($file in $filePaths) {            
+            git -C "${script:llamacppDir}" checkout $file
+        }
+        git -C "${script:llamacppDir}" checkout CMakeLists.txt
+    }
+}
+
+
+# -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+
+
+function build_static() {
+    if ((-not "${env:OLLAMA_SKIP_STATIC_GENERATE}") -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "static"))) {
+        # GCC build for direct linking into the Go binary
+        init_vars
+        # cmake will silently fallback to msvc compilers if mingw isn't in the path, so detect and fail fast
+        # as we need this to be compiled by gcc for golang to be able to link with itx
+        write-host "Checking for MinGW..."
+        # error action ensures we exit on failure
+        get-command gcc
+        get-command mingw32-make
+        $oldTargets = $script:cmakeTargets
+        $script:cmakeTargets = @("llama", "ggml")
+        $script:cmakeDefs = @(
+            "-G", "MinGW Makefiles"
+            "-DCMAKE_C_COMPILER=gcc.exe",
+            "-DCMAKE_CXX_COMPILER=g++.exe",
+            "-DBUILD_SHARED_LIBS=off",
+            "-DLLAMA_NATIVE=off",
+            "-DLLAMA_AVX=off",
+            "-DLLAMA_AVX2=off",
+            "-DLLAMA_AVX512=off",
+            "-DLLAMA_F16C=off",
+            "-DLLAMA_FMA=off")
+        $script:buildDir="../build/windows/${script:ARCH}_static"
+        write-host "Building static library"
+        build
+        $script:cmakeTargets = $oldTargets
+    } else {
+        write-host "Skipping CPU generation step as requested"
+    }
+}
+
+function build_cpu($gen_arch) {
+    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
+        # remaining llama.cpp builds use MSVC 
+        init_vars
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:buildDir="../build/windows/${script:ARCH}/cpu"
+        $script:distDir="$script:DIST_BASE\cpu"
+        write-host "Building LCD CPU"
+        build
+        sign
+        install
+    } else {
+        write-host "Skipping CPU generation step as requested"
+    }
+}
+
+function build_cpu_avx() {
+    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx"))) {
+        init_vars
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
+        $script:distDir="$script:DIST_BASE\cpu_avx"
+        write-host "Building AVX CPU"
+        build
+        sign
+        install
+    } else {
+        write-host "Skipping CPU AVX generation step as requested"
+    }
+}
+
+function build_cpu_avx2() {
+    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx2"))) {
+        init_vars
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
+        $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
+        $script:distDir="$script:DIST_BASE\cpu_avx2"
+        write-host "Building AVX2 CPU"
+        build
+        sign
+        install
+    } else {
+        write-host "Skipping CPU AVX2 generation step as requested"
+    }
+}
+
+function build_cuda() {
+    if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${script:CUDA_LIB_DIR}")) {
+        # Then build cuda as a dynamically loaded library
+        $nvcc = "$script:CUDA_LIB_DIR\nvcc.exe"
+        $script:CUDA_VERSION=(get-item ($nvcc | split-path | split-path)).Basename
+        if ($null -ne $script:CUDA_VERSION) {
+            $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
+        }
+        init_vars
+        $script:buildDir="../build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
+        $script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
+        $script:cmakeDefs += @("-A", "x64", "-DLLAMA_CUDA=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
+        if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
+            write-host "OLLAMA_CUSTOM_CUDA_DEFS=`"${env:OLLAMA_CUSTOM_CUDA_DEFS}`""
+            $script:cmakeDefs +=@("${env:OLLAMA_CUSTOM_CUDA_DEFS}")
+            write-host "building custom CUDA GPU"
+        }
+        build
+        sign
+        install
+
+        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+    } else {
+        write-host "Skipping CUDA generation step"
+    }
+}
+
+function build_oneapi() {
+  if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${env:ONEAPI_ROOT}"))  {
+    # Get oneAPI version
+    $script:ONEAPI_VERSION = icpx --version
+    $script:ONEAPI_VERSION = [regex]::Match($script:ONEAPI_VERSION, '(?<=oneAPI DPC\+\+/C\+\+ Compiler )(?<version>\d+\.\d+\.\d+)').Value
+    if ($null -ne $script:ONEAPI_VERSION) {
+      $script:ONEAPI_VARIANT = "_v" + $script:ONEAPI_VERSION
+    }
+    init_vars
+    $script:buildDir = "../build/windows/${script:ARCH}/oneapi$script:ONEAPI_VARIANT"
+    $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
+    $script:cmakeDefs += @(
+      "-G", "MinGW Makefiles",
+      "-DLLAMA_SYCL=ON",
+      "-DCMAKE_C_COMPILER=icx",
+      "-DCMAKE_CXX_COMPILER=icx",
+      "-DCMAKE_BUILD_TYPE=Release"
+    )
+
+    Write-Host "Building oneAPI"
+    build
+    # Ninja doesn't prefix with config name
+    if ($null -ne $script:DUMPBIN) {
+      & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | Select-String ".dll"
+    }
+    sign
+    install
+
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:distDir}"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:distDir}"
+  } else {
+    Write-Host "Skipping oneAPI generation step"
+  }
+}
+
+function build_rocm() {
+    if ((-not "${env:OLLAMA_SKIP_ROCM_GENERATE}") -and ("${env:HIP_PATH}")) {
+        $script:ROCM_VERSION=(get-item $env:HIP_PATH).Basename
+        if ($null -ne $script:ROCM_VERSION) {
+            $script:ROCM_VARIANT="_v"+$script:ROCM_VERSION
+        }
+
+        init_vars
+        $script:buildDir="../build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
+        $script:distDir="$script:DIST_BASE\rocm$script:ROCM_VARIANT"
+        $script:cmakeDefs += @(
+            "-G", "Ninja", 
+            "-DCMAKE_C_COMPILER=clang.exe",
+            "-DCMAKE_CXX_COMPILER=clang++.exe",
+            "-DLLAMA_HIPBLAS=on",
+            "-DHIP_PLATFORM=amd",
+            "-DLLAMA_AVX=on",
+            "-DLLAMA_AVX2=off",
+            "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
+            "-DAMDGPU_TARGETS=$(amdGPUs)",
+            "-DGPU_TARGETS=$(amdGPUs)"
+            )
+
+        # Make sure the ROCm binary dir is first in the path
+        $env:PATH="$env:HIP_PATH\bin;$env:PATH"
+
+        # We have to clobber the LIB var from the developer shell for clang to work properly
+        $env:LIB=""
+        if ($null -ne $env:OLLAMA_CUSTOM_ROCM_DEFS) {
+            write-host "OLLAMA_CUSTOM_ROCM_DEFS=`"${env:OLLAMA_CUSTOM_ROCM_DEFS}`""
+            $script:cmakeDefs += @("${env:OLLAMA_CUSTOM_ROCM_DEFS}")
+            write-host "building custom ROCM GPU"
+        }
+        write-host "Building ROCm"
+        build
+        # Ninja doesn't prefix with config name
+        ${script:config}=""
+        if ($null -ne $script:DUMPBIN) {
+            & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | select-string ".dll"
+        }
+        sign
+        install
+
+        # Assumes v5.7, may need adjustments for v6
+        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
+        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+        # amdhip64.dll dependency comes from the driver and must be installed on the host to use AMD GPUs
+        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\"
+    } else {
+        write-host "Skipping ROCm generation step"
+    }
+}
+
+init_vars
+if ($($args.count) -eq 0) {
+    git_module_setup
+    apply_patches
+    build_static
+    if ($script:ARCH -eq "arm64") {
+        build_cpu("ARM64")
+    } else { # amd64
+        build_cpu("x64")
+        build_cpu_avx
+        build_cpu_avx2
+        build_cuda
+        build_oneapi
+        build_rocm
+    }
+
+    cleanup
+    write-host "`ngo generate completed.  LLM runners: $(get-childitem -path $script:DIST_BASE)"
+} else {
+    for ( $i = 0; $i -lt $args.count; $i++ ) {
+        write-host "performing $($args[$i])"
+        & $($args[$i])
+    } 
+}
\ No newline at end of file

llm/generate/generate_darwin.go

+package generate
+
+//go:generate bash ./gen_darwin.sh

llm/generate/generate_linux.go

+package generate
+
+//go:generate bash ./gen_linux.sh

llm/generate/generate_windows.go

+package generate
+
+//go:generate powershell -ExecutionPolicy Bypass -File ./gen_windows.ps1

llm/ggla.go

+package llm
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+	"slices"
+)
+
+type containerGGLA struct {
+	version uint32
+}
+
+func (c *containerGGLA) Name() string {
+	return "ggla"
+}
+
+func (c *containerGGLA) Decode(rs io.ReadSeeker) (model, error) {
+	if err := binary.Read(rs, binary.LittleEndian, &c.version); err != nil {
+		return nil, err
+	}
+
+	switch c.version {
+	case 1:
+	default:
+		return nil, errors.New("invalid version")
+	}
+
+	model := newGGLA(c)
+	err := model.decode(rs)
+	return model, err
+}
+
+type ggla struct {
+	*containerGGLA
+
+	kv      KV
+	tensors []*Tensor
+}
+
+func newGGLA(container *containerGGLA) *ggla {
+	return &ggla{
+		containerGGLA: container,
+		kv:            make(KV),
+	}
+}
+
+func (llm *ggla) KV() KV {
+	return llm.kv
+}
+
+func (llm *ggla) Tensors() Tensors {
+	return llm.tensors
+}
+
+func (llm *ggla) decode(rs io.ReadSeeker) error {
+	var r uint32
+	if err := binary.Read(rs, binary.LittleEndian, &r); err != nil {
+		return err
+	}
+	llm.kv["r"] = r
+
+	var alpha uint32
+	if err := binary.Read(rs, binary.LittleEndian, &alpha); err != nil {
+		return err
+	}
+	llm.kv["alpha"] = alpha
+
+	for {
+		var dims uint32
+		if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {
+			return err
+		}
+
+		var namesize uint32
+		if err := binary.Read(rs, binary.LittleEndian, &namesize); err != nil {
+			return err
+		}
+
+		var t Tensor
+		if err := binary.Read(rs, binary.LittleEndian, &t.Kind); err != nil {
+			return err
+		}
+
+		t.Shape = make([]uint64, dims)
+		for i := 0; uint32(i) < dims; i++ {
+			var shape32 uint32
+			if err := binary.Read(rs, binary.LittleEndian, &shape32); err != nil {
+				return err
+			}
+
+			t.Shape[i] = uint64(shape32)
+		}
+
+		// ggla tensor shape is reversed
+		// ref: https://github.com/ggerganov/llama.cpp/blob/29ae62d2ae163e2b68aa0ad3bf2ab4636de0c957/convert-lora-to-ggml.py#L44
+		slices.Reverse(t.Shape)
+
+		name := make([]byte, namesize)
+		if err := binary.Read(rs, binary.LittleEndian, &name); err != nil {
+			return err
+		}
+
+		t.Name = string(name)
+
+		offset, err := rs.Seek(0, io.SeekCurrent)
+		if err != nil {
+			return err
+		}
+
+		if _, err := rs.Seek((offset+31)&-32, io.SeekStart); err != nil {
+			return err
+		}
+
+		offset, err = rs.Seek(0, io.SeekCurrent)
+		if err != nil {
+			return err
+		}
+
+		t.Offset = uint64(offset)
+
+		if _, err := rs.Seek(int64(t.Size()), io.SeekCurrent); err != nil {
+			return err
+		}
+
+		llm.tensors = append(llm.tensors, &t)
+	}
+}

llm/ggml.go

+package llm
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"strings"
+)
+
+type GGML struct {
+	container
+	model
+}
+
+type model interface {
+	KV() KV
+	Tensors() Tensors
+}
+
+type KV map[string]any
+
+func (kv KV) u64(key string) uint64 {
+	switch v := kv[key].(type) {
+	case uint64:
+		return v
+	case uint32:
+		return uint64(v)
+	case float64:
+		return uint64(v)
+	default:
+		return 0
+	}
+}
+
+func (kv KV) Architecture() string {
+	if s, ok := kv["general.architecture"].(string); ok {
+		return s
+	}
+
+	return "unknown"
+}
+
+func (kv KV) ParameterCount() uint64 {
+	return kv.u64("general.parameter_count")
+}
+
+func (kv KV) FileType() fileType {
+	if u64 := kv.u64("general.file_type"); u64 > 0 {
+		return fileType(uint32(u64))
+	}
+
+	return fileTypeUnknown
+}
+
+func (kv KV) BlockCount() uint64 {
+	return kv.u64(fmt.Sprintf("%s.block_count", kv.Architecture()))
+}
+
+func (kv KV) HeadCount() uint64 {
+	return kv.u64(fmt.Sprintf("%s.attention.head_count", kv.Architecture()))
+}
+
+func (kv KV) HeadCountKV() uint64 {
+	if headCountKV := kv.u64(fmt.Sprintf("%s.attention.head_count_kv", kv.Architecture())); headCountKV > 0 {
+		return headCountKV
+	}
+
+	return 1
+}
+
+func (kv KV) GQA() uint64 {
+	return kv.HeadCount() / kv.HeadCountKV()
+}
+
+func (kv KV) EmbeddingLength() uint64 {
+	return kv.u64(fmt.Sprintf("%s.embedding_length", kv.Architecture()))
+}
+
+func (kv KV) ContextLength() uint64 {
+	return kv.u64(fmt.Sprintf("%s.context_length", kv.Architecture()))
+}
+
+type Tensors []*Tensor
+
+func (ts Tensors) Layers() map[string]Layer {
+	layers := make(map[string]Layer)
+	for _, t := range ts {
+		parts := strings.Split(t.Name, ".")
+		if parts[0] == "blk" {
+			// join first and second part, e.g. blk.%d
+			parts = append([]string{fmt.Sprintf("%s.%s", parts[0], parts[1])}, parts[2:]...)
+		}
+
+		if _, ok := layers[parts[0]]; !ok {
+			layers[parts[0]] = make(Layer)
+		}
+
+		layers[parts[0]][strings.Join(parts[1:], ".")] = t
+	}
+
+	return layers
+}
+
+type Layer map[string]*Tensor
+
+func (l Layer) size() (size uint64) {
+	for _, t := range l {
+		size += t.Size()
+	}
+
+	return size
+}
+
+type Tensor struct {
+	Name   string `json:"name"`
+	Kind   uint32 `json:"kind"`
+	Offset uint64 `json:"-"`
+
+	// Shape is the number of elements in each dimension
+	Shape []uint64 `json:"shape"`
+
+	io.WriterTo `json:"-"`
+}
+
+func (t Tensor) blockSize() uint64 {
+	switch t.Kind {
+	case 0, 1, 24, 25, 26, 27, 28, 30: // F32, F16, I8, I16, I32, I64, F64, BF16
+		return 1
+	case 2, 3, 4, 5, 6, 7, 8, 9, 20: // Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, Q8_1, IQ4_NL
+		return 32
+	default: // All others
+		return 256
+	}
+}
+
+func (t Tensor) typeSize() uint64 {
+	blockSize := t.blockSize()
+
+	switch t.Kind {
+	case 0: // FP32
+		return 4
+	case 1: // FP16
+		return 2
+	case 2: // Q4_0
+		return 2 + blockSize/2
+	case 3: // Q4_1
+		return 2 + 2 + blockSize/2
+	case 6: // Q5_0
+		return 2 + 4 + blockSize/2
+	case 7: // Q5_1
+		return 2 + 2 + 4 + blockSize/2
+	case 8: // Q8_0
+		return 2 + blockSize
+	case 9: // Q8_1
+		return 4 + 4 + blockSize
+	case 10: // Q2_K
+		return blockSize/16 + blockSize/4 + 2 + 2
+	case 11: // Q3_K
+		return blockSize/8 + blockSize/4 + 12 + 2
+	case 12: // Q4_K
+		return 2 + 2 + 12 + blockSize/2
+	case 13: // Q5_K
+		return 2 + 2 + 12 + blockSize/8 + blockSize/2
+	case 14: // Q6_K
+		return blockSize/2 + blockSize/4 + blockSize/16 + 2
+	case 15: // Q8_K
+		return 2 + blockSize + 2*blockSize/16
+	case 16: // IQ2_XXS
+		return 2 + 2*blockSize/8
+	case 17: // IQ2_XS
+		return 2 + 2*blockSize/8 + blockSize/32
+	case 18: // IQ3_XXS
+		return 2 + blockSize/4 + blockSize/8
+	case 19: // IQ1_S
+		return 2 + blockSize/8 + blockSize/16
+	case 20: // IQ4_NL
+		return 2 + blockSize/2
+	case 21: // IQ3_S
+		return 2 + blockSize/4 + blockSize/8 + blockSize/32 + 4
+	case 22: // IQ2_S
+		return 2 + blockSize/4 + blockSize/16
+	case 23: // IQ4_XS
+		return 2 + 2 + blockSize/2 + blockSize/64
+	case 24: // I8
+		return 1
+	case 25: // I16
+		return 2
+	case 26: // I32
+		return 4
+	case 27: // I64
+		return 8
+	case 28: // F64
+		return 8
+	case 29: // IQ1_M
+		return blockSize/8 + blockSize/16 + blockSize/32
+	default:
+		return 0
+	}
+}
+
+func (t Tensor) parameters() uint64 {
+	var count uint64 = 1
+	for _, n := range t.Shape {
+		count *= n
+	}
+	return count
+}
+
+func (t Tensor) Size() uint64 {
+	return t.parameters() * t.typeSize() / t.blockSize()
+}
+
+type container interface {
+	Name() string
+	Decode(io.ReadSeeker) (model, error)
+}
+
+const (
+	// Magic constant for `ggml` files (unversioned).
+	FILE_MAGIC_GGML = 0x67676d6c
+	// Magic constant for `ggml` files (versioned, ggmf).
+	FILE_MAGIC_GGMF = 0x67676d66
+	// Magic constant for `ggml` files (versioned, ggjt).
+	FILE_MAGIC_GGJT = 0x67676a74
+	// Magic constant for `ggla` files (LoRA adapter).
+	FILE_MAGIC_GGLA = 0x67676C61
+	// Magic constant for `gguf` files (versioned, gguf)
+	FILE_MAGIC_GGUF_LE = 0x46554747
+	FILE_MAGIC_GGUF_BE = 0x47475546
+)
+
+var ErrUnsupportedFormat = errors.New("unsupported model format")
+
+func DetectGGMLType(b []byte) string {
+	switch binary.LittleEndian.Uint32(b[:4]) {
+	case FILE_MAGIC_GGML:
+		return "ggml"
+	case FILE_MAGIC_GGMF:
+		return "ggmf"
+	case FILE_MAGIC_GGJT:
+		return "ggjt"
+	case FILE_MAGIC_GGLA:
+		return "ggla"
+	case FILE_MAGIC_GGUF_LE, FILE_MAGIC_GGUF_BE:
+		return "gguf"
+	default:
+		return ""
+	}
+}
+
+func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
+	var magic uint32
+	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {
+		return nil, 0, err
+	}
+
+	var c container
+	switch magic {
+	case FILE_MAGIC_GGML, FILE_MAGIC_GGMF, FILE_MAGIC_GGJT:
+		return nil, 0, ErrUnsupportedFormat
+	case FILE_MAGIC_GGLA:
+		c = &containerGGLA{}
+	case FILE_MAGIC_GGUF_LE:
+		c = &containerGGUF{ByteOrder: binary.LittleEndian}
+	case FILE_MAGIC_GGUF_BE:
+		c = &containerGGUF{ByteOrder: binary.BigEndian}
+	default:
+		return nil, 0, errors.New("invalid file magic")
+	}
+
+	model, err := c.Decode(rs)
+	if errors.Is(err, io.EOF) {
+		// noop
+	} else if err != nil {
+		return nil, 0, err
+	}
+
+	offset, err := rs.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return nil, 0, err
+	}
+
+	// final model type
+	return &GGML{
+		container: c,
+		model:     model,
+	}, offset, nil
+}
+
+func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload uint64) {
+	embedding := llm.KV().EmbeddingLength()
+	heads := llm.KV().HeadCount()
+	headsKV := llm.KV().HeadCountKV()
+	vocab := uint64(len(llm.KV()["tokenizer.ggml.tokens"].([]any)))
+
+	layers := llm.Tensors().Layers()
+
+	switch llm.KV().Architecture() {
+	case "llama":
+		fullOffload = 4 * batch * (1 + 4*embedding + context*(1+heads))
+
+		partialOffload = 4 * batch * embedding
+		partialOffload += max(
+			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embedding/heads*headsKV),
+			4*batch*(embedding+vocab)+embedding*vocab*105/128,
+		)
+
+		if ffnGateExpsWeight, ok := layers["blk.0"]["ffn_gate_exps.weight"]; ok {
+			// mixtral 8x22b
+			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
+			partialOffload = max(
+				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
+				4*(context*batch*heads+context*embedding/heads*headsKV+batch*1024+embedding/heads*headsKV*batch),
+			)
+		} else if ffnGateWeight, ok := layers["blk.0"]["ffn_gate.0.weight"]; ok {
+			// mixtral 8x7b
+			ffnGateWeight1 := ffnGateWeight.Shape[1]
+			fullOffload = 4 * batch * (2 + 3*embedding + context*(1+heads) + 2*headsKV + ffnGateWeight1)
+			partialOffload = max(
+				4*batch*(3+embedding/heads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
+				4*batch*(1+2*embedding+context*(1+heads))+embedding*(6*context*headsKV/heads+embedding*9/16),
+			)
+		}
+	case "gemma":
+		fullOffload = 4 * batch * (embedding + vocab)
+		partialOffload = 4*batch*(2*embedding+vocab+1) + embedding*vocab*105/128
+	case "command-r":
+		fullOffload = max(
+			4*batch*(embedding+vocab),
+			4*batch*(2+4*embedding+context*(1+heads)),
+		)
+
+		partialOffload = max(
+			4*batch*(embedding+vocab)+embedding*vocab*105/128,
+			4*batch*(1+2*embedding+context*(1+heads))+4*embedding*context+embedding*embedding*9/16,
+		)
+	case "qwen2":
+		fullOffload = max(
+			4*batch*(embedding+vocab),
+			4*batch*(1+2*embedding+context+context*heads),
+		)
+
+		partialOffload = max(
+			4*batch*(embedding+vocab)+embedding*vocab*105/128,
+			4*(batch*(1+2*embedding+context*(1+heads))+embedding*(1+context)),
+		)
+	case "phi2":
+		fullOffload = max(
+			4*batch*(embedding+vocab),
+			4*batch*(1+4*embedding+context+context*heads),
+		)
+
+		partialOffload = max(
+			4*batch*(2*embedding+vocab)+embedding*vocab*105/128,
+			4*batch*(2+3*embedding+context+context*heads),
+		)
+	case "stablelm":
+		fullOffload = 4 * batch * (context*(1+heads) + 3*embedding + 2)
+		partialOffload = max(
+			4*batch*(vocab+2*embedding),
+			fullOffload,
+		)
+	}
+
+	return
+}

llm/gguf.go

+package llm
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"io"
+	"strings"
+
+	"log/slog"
+)
+
+type containerGGUF struct {
+	ByteOrder binary.ByteOrder
+
+	Version uint32
+
+	V1 struct {
+		NumTensor uint32
+		NumKV     uint32
+	}
+
+	V2 struct {
+		NumTensor uint64
+		NumKV     uint64
+	}
+
+	V3 struct {
+		NumTensor uint64
+		NumKV     uint64
+	}
+}
+
+func (c *containerGGUF) Name() string {
+	return "gguf"
+}
+
+func (c *containerGGUF) Decode(rs io.ReadSeeker) (model, error) {
+	if err := binary.Read(rs, c.ByteOrder, &c.Version); err != nil {
+		return nil, err
+	}
+
+	var err error
+	switch c.Version {
+	case 1:
+		err = binary.Read(rs, c.ByteOrder, &c.V1)
+	case 2:
+		err = binary.Read(rs, c.ByteOrder, &c.V2)
+	default:
+		err = binary.Read(rs, c.ByteOrder, &c.V3)
+	}
+	if err != nil {
+		return nil, err
+	}
+
+	model := newGGUF(c)
+	slog.Debug(fmt.Sprintf("model = %#v", model))
+	if err := model.Decode(rs); err != nil {
+		return nil, err
+	}
+
+	return model, nil
+}
+
+const (
+	ggufTypeUint8 uint32 = iota
+	ggufTypeInt8
+	ggufTypeUint16
+	ggufTypeInt16
+	ggufTypeUint32
+	ggufTypeInt32
+	ggufTypeFloat32
+	ggufTypeBool
+	ggufTypeString
+	ggufTypeArray
+	ggufTypeUint64
+	ggufTypeInt64
+	ggufTypeFloat64
+)
+
+type gguf struct {
+	*containerGGUF
+
+	kv      KV
+	tensors []*Tensor
+
+	parameters uint64
+}
+
+func newGGUF(container *containerGGUF) *gguf {
+	return &gguf{
+		containerGGUF: container,
+		kv:            make(KV),
+	}
+}
+
+func NewGGUFV3(bo binary.ByteOrder) *gguf {
+	return newGGUF(&containerGGUF{ByteOrder: bo, Version: 3})
+}
+
+func (llm *gguf) KV() KV {
+	return llm.kv
+}
+
+func (llm *gguf) Tensors() Tensors {
+	return llm.tensors
+}
+
+func (llm *gguf) numTensor() uint64 {
+	switch llm.Version {
+	case 1:
+		return uint64(llm.V1.NumTensor)
+	case 2:
+		return llm.V2.NumTensor
+	default:
+		return llm.V3.NumTensor
+	}
+}
+
+func (llm *gguf) numKV() uint64 {
+	switch llm.Version {
+	case 1:
+		return uint64(llm.V1.NumKV)
+	case 2:
+		return llm.V2.NumKV
+	default:
+		return llm.V3.NumKV
+	}
+}
+
+func (llm *gguf) Decode(rs io.ReadSeeker) error {
+	// decode key-values
+	for i := 0; uint64(i) < llm.numKV(); i++ {
+		k, err := readGGUFString(llm, rs)
+		if err != nil {
+			return err
+		}
+
+		t, err := readGGUF[uint32](llm, rs)
+		if err != nil {
+			return err
+		}
+
+		var v any
+		switch t {
+		case ggufTypeUint8:
+			v, err = readGGUF[uint8](llm, rs)
+		case ggufTypeInt8:
+			v, err = readGGUF[int8](llm, rs)
+		case ggufTypeUint16:
+			v, err = readGGUF[uint16](llm, rs)
+		case ggufTypeInt16:
+			v, err = readGGUF[int16](llm, rs)
+		case ggufTypeUint32:
+			v, err = readGGUF[uint32](llm, rs)
+		case ggufTypeInt32:
+			v, err = readGGUF[int32](llm, rs)
+		case ggufTypeUint64:
+			v, err = readGGUF[uint64](llm, rs)
+		case ggufTypeInt64:
+			v, err = readGGUF[int64](llm, rs)
+		case ggufTypeFloat32:
+			v, err = readGGUF[float32](llm, rs)
+		case ggufTypeFloat64:
+			v, err = readGGUF[float64](llm, rs)
+		case ggufTypeBool:
+			v, err = readGGUF[bool](llm, rs)
+		case ggufTypeString:
+			v, err = readGGUFString(llm, rs)
+		case ggufTypeArray:
+			v, err = readGGUFArray(llm, rs)
+		default:
+			return fmt.Errorf("invalid type: %d", t)
+		}
+
+		if err != nil {
+			return err
+		}
+
+		llm.kv[k] = v
+	}
+
+	// decode tensors
+	for i := 0; uint64(i) < llm.numTensor(); i++ {
+		name, err := readGGUFString(llm, rs)
+		if err != nil {
+			return err
+		}
+
+		// dims is the number of dimensions in the tensor
+		dims, err := readGGUF[uint32](llm, rs)
+		if err != nil {
+			return err
+		}
+
+		shape := [4]uint64{1, 1, 1, 1}
+		for i := 0; uint32(i) < dims; i++ {
+			shape[i], err = readGGUF[uint64](llm, rs)
+			if err != nil {
+				return err
+			}
+		}
+
+		kind, err := readGGUF[uint32](llm, rs)
+		if err != nil {
+			return err
+		}
+
+		offset, err := readGGUF[uint64](llm, rs)
+		if err != nil {
+			return err
+		}
+
+		tensor := Tensor{
+			Name:   name,
+			Kind:   kind,
+			Offset: offset,
+			Shape:  shape[:],
+		}
+
+		llm.tensors = append(llm.tensors, &tensor)
+		llm.parameters += tensor.parameters()
+	}
+
+	// patch KV with parameter count
+	llm.kv["general.parameter_count"] = llm.parameters
+
+	alignment, ok := llm.kv["general.alignment"].(uint32)
+	if !ok {
+		alignment = 32
+	}
+
+	offset, err := rs.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+
+	padding := llm.padding(offset, int64(alignment))
+	if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
+		return err
+	}
+
+	for _, tensor := range llm.tensors {
+		if _, err := rs.Seek(int64(tensor.Size()), io.SeekCurrent); err != nil {
+			return err
+		}
+
+		padding := llm.padding(int64(tensor.Size()), int64(alignment))
+		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func readGGUF[T any](llm *gguf, r io.Reader) (T, error) {
+	var t T
+	err := binary.Read(r, llm.ByteOrder, &t)
+	return t, err
+}
+
+func writeGGUF[V any](llm *gguf, w io.Writer, t uint32, v V) error {
+	if err := binary.Write(w, llm.ByteOrder, t); err != nil {
+		return err
+	}
+
+	return binary.Write(w, llm.ByteOrder, v)
+}
+
+func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
+	var length uint64
+	if err := binary.Read(r, llm.ByteOrder, &length); err != nil {
+		return "", err
+	}
+
+	var b bytes.Buffer
+	if _, err := io.CopyN(&b, r, int64(length)); err != nil {
+		return "", err
+	}
+
+	// gguf v1 strings are null-terminated
+	b.Truncate(b.Len() - 1)
+
+	return b.String(), nil
+}
+
+func readGGUFString(llm *gguf, r io.Reader) (string, error) {
+	if llm.Version == 1 {
+		return readGGUFV1String(llm, r)
+	}
+
+	var length uint64
+	if err := binary.Read(r, llm.ByteOrder, &length); err != nil {
+		return "", err
+	}
+
+	var b bytes.Buffer
+	if _, err := io.CopyN(&b, r, int64(length)); err != nil {
+		return "", err
+	}
+
+	return b.String(), nil
+}
+
+func writeGGUFString(llm *gguf, w io.Writer, s string) error {
+	if err := binary.Write(w, llm.ByteOrder, ggufTypeString); err != nil {
+		return err
+	}
+
+	if err := binary.Write(w, llm.ByteOrder, uint64(len(s))); err != nil {
+		return err
+	}
+
+	_, err := io.Copy(w, strings.NewReader(s))
+	return err
+}
+
+func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
+	t, err := readGGUF[uint32](llm, r)
+	if err != nil {
+		return nil, err
+	}
+
+	n, err := readGGUF[uint32](llm, r)
+	if err != nil {
+		return nil, err
+	}
+
+	for i := 0; uint32(i) < n; i++ {
+		var e any
+		switch t {
+		case ggufTypeUint8:
+			e, err = readGGUF[uint8](llm, r)
+		case ggufTypeInt8:
+			e, err = readGGUF[int8](llm, r)
+		case ggufTypeUint16:
+			e, err = readGGUF[uint16](llm, r)
+		case ggufTypeInt16:
+			e, err = readGGUF[int16](llm, r)
+		case ggufTypeUint32:
+			e, err = readGGUF[uint32](llm, r)
+		case ggufTypeInt32:
+			e, err = readGGUF[int32](llm, r)
+		case ggufTypeUint64:
+			e, err = readGGUF[uint64](llm, r)
+		case ggufTypeInt64:
+			e, err = readGGUF[int64](llm, r)
+		case ggufTypeFloat32:
+			e, err = readGGUF[float32](llm, r)
+		case ggufTypeFloat64:
+			e, err = readGGUF[float64](llm, r)
+		case ggufTypeBool:
+			e, err = readGGUF[bool](llm, r)
+		case ggufTypeString:
+			e, err = readGGUFV1String(llm, r)
+		default:
+			return nil, fmt.Errorf("invalid array type: %d", t)
+		}
+		if err != nil {
+			return nil, err
+		}
+
+		a = append(a, e)
+	}
+
+	return
+}
+
+func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
+	if llm.Version == 1 {
+		return readGGUFV1Array(llm, r)
+	}
+
+	t, err := readGGUF[uint32](llm, r)
+	if err != nil {
+		return nil, err
+	}
+
+	n, err := readGGUF[uint64](llm, r)
+	if err != nil {
+		return nil, err
+	}
+
+	for i := 0; uint64(i) < n; i++ {
+		var e any
+		switch t {
+		case ggufTypeUint8:
+			e, err = readGGUF[uint8](llm, r)
+		case ggufTypeInt8:
+			e, err = readGGUF[int8](llm, r)
+		case ggufTypeUint16:
+			e, err = readGGUF[uint16](llm, r)
+		case ggufTypeInt16:
+			e, err = readGGUF[int16](llm, r)
+		case ggufTypeUint32:
+			e, err = readGGUF[uint32](llm, r)
+		case ggufTypeInt32:
+			e, err = readGGUF[int32](llm, r)
+		case ggufTypeUint64:
+			e, err = readGGUF[uint64](llm, r)
+		case ggufTypeInt64:
+			e, err = readGGUF[int64](llm, r)
+		case ggufTypeFloat32:
+			e, err = readGGUF[float32](llm, r)
+		case ggufTypeFloat64:
+			e, err = readGGUF[float64](llm, r)
+		case ggufTypeBool:
+			e, err = readGGUF[bool](llm, r)
+		case ggufTypeString:
+			e, err = readGGUFString(llm, r)
+		default:
+			return nil, fmt.Errorf("invalid array type: %d", t)
+		}
+		if err != nil {
+			return nil, err
+		}
+
+		a = append(a, e)
+	}
+
+	return
+}
+
+func writeGGUFArray[S ~[]E, E any](llm *gguf, w io.Writer, t uint32, s S) error {
+	if err := binary.Write(w, llm.ByteOrder, ggufTypeArray); err != nil {
+		return err
+	}
+
+	if err := binary.Write(w, llm.ByteOrder, t); err != nil {
+		return err
+	}
+
+	if err := binary.Write(w, llm.ByteOrder, uint64(len(s))); err != nil {
+		return err
+	}
+
+	for _, e := range s {
+		if err := binary.Write(w, llm.ByteOrder, e); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+var ggufKVOrder = map[string][]string{
+	"llama": {
+		"general.architecture",
+		"general.name",
+		"llama.vocab_size",
+		"llama.context_length",
+		"llama.embedding_length",
+		"llama.block_count",
+		"llama.feed_forward_length",
+		"llama.attention.head_count",
+		"llama.attention.head_count_kv",
+		"llama.attention.layer_norm_rms_epsilon",
+		"llama.rope.freq_base",
+		"llama.rope.dimension_count",
+		"llama.expert_count",
+		"llama.expert_used_count",
+		"gemma.context_length",
+		"gemma.embedding_length",
+		"gemma.block_count",
+		"gemma.feed_forward_length",
+		"gemma.attention.head_count",
+		"gemma.attention.head_count_kv",
+		"gemma.attention.layer_norm_rms_epsilon",
+		"gemma.attention.key_length",
+		"gemma.attention.value_length",
+		"general.file_type",
+		"tokenizer.ggml.pre",
+		"tokenizer.ggml.model",
+		"tokenizer.ggml.tokens",
+		"tokenizer.ggml.scores",
+		"tokenizer.ggml.merges",
+		"tokenizer.ggml.token_type",
+		"tokenizer.ggml.bos_token_id",
+		"tokenizer.ggml.eos_token_id",
+		"tokenizer.ggml.unknown_token_id",
+		"tokenizer.ggml.padding_token_id",
+		"tokenizer.ggml.add_bos_token",
+		"tokenizer.ggml.add_eos_token",
+		"tokenizer.chat_template",
+	},
+}
+
+func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
+	switch llm.Version {
+	case 3:
+		llm.V3.NumTensor = uint64(len(tensors))
+		llm.V3.NumKV = uint64(len(kv))
+	default:
+		return fmt.Errorf("not implemented: ggufv%d", llm.Version)
+	}
+
+	if err := binary.Write(ws, llm.ByteOrder, []byte("GGUF")); err != nil {
+		return err
+	}
+
+	if err := binary.Write(ws, llm.ByteOrder, llm.Version); err != nil {
+		return err
+	}
+
+	if err := binary.Write(ws, llm.ByteOrder, llm.numTensor()); err != nil {
+		return err
+	}
+
+	if err := binary.Write(ws, llm.ByteOrder, llm.numKV()); err != nil {
+		return err
+	}
+
+	kvCheck := make(map[string]bool)
+	for k := range kv {
+		kvCheck[k] = false
+	}
+
+	for _, k := range ggufKVOrder["llama"] {
+		v, ok := kv[k]
+		if !ok {
+			continue
+		}
+		kvCheck[k] = true
+
+		if err := binary.Write(ws, llm.ByteOrder, uint64(len(k))); err != nil {
+			return err
+		}
+
+		if err := binary.Write(ws, llm.ByteOrder, []byte(k)); err != nil {
+			return err
+		}
+
+		var err error
+		switch v := v.(type) {
+		case uint32:
+			err = writeGGUF(llm, ws, ggufTypeUint32, v)
+		case float32:
+			err = writeGGUF(llm, ws, ggufTypeFloat32, v)
+		case bool:
+			err = writeGGUF(llm, ws, ggufTypeBool, v)
+		case string:
+			err = writeGGUFString(llm, ws, v)
+		case []int32:
+			err = writeGGUFArray(llm, ws, ggufTypeInt32, v)
+		case []uint32:
+			err = writeGGUFArray(llm, ws, ggufTypeUint32, v)
+		case []float32:
+			err = writeGGUFArray(llm, ws, ggufTypeFloat32, v)
+		case []string:
+			if err := binary.Write(ws, llm.ByteOrder, ggufTypeArray); err != nil {
+				return err
+			}
+
+			if err := binary.Write(ws, llm.ByteOrder, ggufTypeString); err != nil {
+				return err
+			}
+
+			if err := binary.Write(ws, llm.ByteOrder, uint64(len(v))); err != nil {
+				return err
+			}
+
+			for _, e := range v {
+				if err := binary.Write(ws, llm.ByteOrder, uint64(len(e))); err != nil {
+					return err
+				}
+
+				if err := binary.Write(ws, llm.ByteOrder, []byte(e)); err != nil {
+					return err
+				}
+			}
+		default:
+			return fmt.Errorf("improper type for '%s'", k)
+		}
+		if err != nil {
+			return err
+		}
+	}
+
+	for k, v := range kvCheck {
+		if !v {
+			return fmt.Errorf("Didn't know how to write kv %s", k)
+		}
+	}
+
+	for _, tensor := range tensors {
+		if err := binary.Write(ws, llm.ByteOrder, uint64(len(tensor.Name))); err != nil {
+			return err
+		}
+
+		if err := binary.Write(ws, llm.ByteOrder, []byte(tensor.Name)); err != nil {
+			return err
+		}
+
+		dims := 0
+		for cnt := 0; cnt < len(tensor.Shape); cnt++ {
+			if tensor.Shape[cnt] > 0 {
+				dims++
+			}
+		}
+
+		if err := binary.Write(ws, llm.ByteOrder, uint32(dims)); err != nil {
+			return err
+		}
+
+		for i := 0; i < dims; i++ {
+			if err := binary.Write(ws, llm.ByteOrder, uint64(tensor.Shape[dims-1-i])); err != nil {
+				return err
+			}
+		}
+
+		if err := binary.Write(ws, llm.ByteOrder, tensor.Kind); err != nil {
+			return err
+		}
+
+		if err := binary.Write(ws, llm.ByteOrder, tensor.Offset); err != nil {
+			return err
+		}
+	}
+
+	offset, err := ws.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+
+	var alignment int64 = 32
+	padding := llm.padding(offset, alignment)
+	if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
+		return err
+	}
+
+	for _, tensor := range tensors {
+		if _, err := tensor.WriteTo(ws); err != nil {
+			return err
+		}
+
+		offset, err := ws.Seek(0, io.SeekCurrent)
+		if err != nil {
+			return err
+		}
+
+		padding := llm.padding(offset, alignment)
+		if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (gguf) padding(offset, align int64) int64 {
+	return (align - offset%align) % align
+}

llm/llama.cpp/.devops/cloud-v-pipeline

+node('x86_runner1'){            // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries
+    stage('Cleanup'){
+        cleanWs()               // Cleaning previous CI build in workspace
+    }
+    stage('checkout repo'){
+        retry(5){               // Retry if the cloning fails due to some reason
+            checkout scm        // Clone the repo on Runner
+        }
+    }
+    stage('Compiling llama.cpp'){
+        sh'''#!/bin/bash
+            make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V
+        '''
+    }
+    stage('Running llama.cpp'){
+        sh'''#!/bin/bash
+            module load gnu-bin2/0.1            # loading latest versions of vector qemu and vector gcc
+            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./main -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
+            cat llama_log.txt                   # Printing results
+        '''
+    }
+}

llm/llama.cpp/.devops/full-cuda.Dockerfile

+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable CUDA
+ENV LLAMA_CUDA=1
+# Enable cURL
+ENV LLAMA_CURL=1
+
+RUN make
+
+ENTRYPOINT ["/app/.devops/tools.sh"]

llm/llama.cpp/.devops/full-rocm.Dockerfile

+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+# Enable cURL
+ENV LLAMA_CURL=1
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
+RUN make
+
+ENTRYPOINT ["/app/.devops/tools.sh"]

llm/llama.cpp/.devops/full.Dockerfile

+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+ENV LLAMA_CURL=1
+
+
+RUN make
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT ["/app/.devops/tools.sh"]

llm/llama.cpp/.devops/llama-cpp-clblast.srpm.spec

+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://docs.fedoraproject.org/en-US/quick-docs/creating-rpm-packages
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp-clblast
+Version:        %( date "+%%Y%%m%%d" )
+Release:        1%{?dist}
+Summary:        OpenCL Inference of LLaMA model in C/C++
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git mesa-libOpenCL-devel clblast-devel
+Requires:       clblast
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j LLAMA_CLBLAST=1
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p main %{buildroot}%{_bindir}/llamaclblast
+cp -p server %{buildroot}%{_bindir}/llamaclblastserver
+cp -p simple %{buildroot}%{_bindir}/llamaclblastsimple
+
+mkdir -p %{buildroot}/usr/lib/systemd/system
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamaclblast.service
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llamaclblastserver $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamaclblast
+%{_bindir}/llamaclblastserver
+%{_bindir}/llamaclblastsimple
+/usr/lib/systemd/system/llamaclblast.service
+%config /etc/sysconfig/llama
+
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

llm/llama.cpp/.devops/llama-cpp-cuda.srpm.spec

+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://docs.fedoraproject.org/en-US/quick-docs/creating-rpm-packages
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp-cuda
+Version:        %( date "+%%Y%%m%%d" )
+Release:        1%{?dist}
+Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git cuda-toolkit
+Requires:       cuda-toolkit
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j LLAMA_CUDA=1
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p main %{buildroot}%{_bindir}/llamacppcuda
+cp -p server %{buildroot}%{_bindir}/llamacppcudaserver
+cp -p simple %{buildroot}%{_bindir}/llamacppcudasimple
+
+mkdir -p %{buildroot}/usr/lib/systemd/system
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacuda.service
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llamacppcudaserver $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamacppcuda
+%{_bindir}/llamacppcudaserver
+%{_bindir}/llamacppcudasimple
+/usr/lib/systemd/system/llamacuda.service
+%config /etc/sysconfig/llama
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

llm/llama.cpp/.devops/llama-cpp.srpm.spec

+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://docs.fedoraproject.org/en-US/quick-docs/creating-rpm-packages
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+#    In the meantime, YYYYMMDD format will be used.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp
+Version:        %( date "+%%Y%%m%%d" )
+Release:        1%{?dist}
+Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git libstdc++-devel
+Requires:       libstdc++
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+Models are not included in this package and must be downloaded separately.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p main %{buildroot}%{_bindir}/llama
+cp -p server %{buildroot}%{_bindir}/llamaserver
+cp -p simple %{buildroot}%{_bindir}/llamasimple
+
+mkdir -p %{buildroot}/usr/lib/systemd/system
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llama.service
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llamaserver $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llama
+%{_bindir}/llamaserver
+%{_bindir}/llamasimple
+/usr/lib/systemd/system/llama.service
+%config /etc/sysconfig/llama
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

llm/llama.cpp/.devops/main-cuda.Dockerfile

+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable CUDA
+ENV LLAMA_CUDA=1
+
+RUN make
+
+FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
+
+COPY --from=build /app/main /main
+
+ENTRYPOINT [ "/main" ]

llm/llama.cpp/.devops/main-intel.Dockerfile

+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
+
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
+    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
+    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
+    rm /etc/apt/sources.list.d/intel-graphics.list && \
+    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
+    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
+    chmod 644 /usr/share/keyrings/intel-graphics.gpg
+
+ARG LLAMA_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
+        echo "LLAMA_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
+    fi && \
+    cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake --build build --config Release --target main
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
+
+COPY --from=build /app/build/bin/main /main
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/main" ]

llm/llama.cpp/.devops/main-rocm.Dockerfile

+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT [ "/app/main" ]